Abstract

Data structures for non-volatile memories have to be designed such that they can be atomically modified using transactions. Existing atomicity methods require data to be copied in the critical path which significantly increases the latency of transactions. These overheads are further amplified for transactions on byte-addressable persistent memories where often the byte ranges modified for data structure updates are significantly smaller compared to the granularity at which data can be efficiently copied and logged. We propose Kamino-Tx that provides a new way to perform transactional updates on non-volatile byte-addressable memories (NVM) without requiring any copying of data in the critical path. Kamino-Tx maintains an additional copy of data off the critical path to achieve atomicity. But in doing so Kamino-Tx has to overcome two important challenges of safety and minimizing NVM storage overhead. We propose a more dynamic approach to maintaining the additional copy of data to reduce storage overheads. To further mitigate the storage overhead of using Kamino-Tx in a replicated setting, we develop Kamino-Tx-Chain, a variant of Chain Replication where replicas perform in-place updates and do not maintain data copies locally; replicas in Kamino-Tx-Chain leverage other replicas as copies to roll back or forward for atomicity. Our results show that using Kamino-Tx increases throughput by up to 9.5x for unreplicated systems and up to 2.2x for replicated settings.